Evaluation of a Hybrid Approach to Legal Entity Recognition

Abstrak

The Brazilian judiciary system’s extensive volume of documents and technical language necessitates efficient methods for automating legal text analysis, where Legal Entity Recognition (LER) presents a significant challenge. This study evaluated the performance of LER models within the Brazilian legal domain through a comprehensive assessment across all publicly available Portuguese legal datasets: LeNER-BR, CDJur-BR, and UlyssesNER-BR. Eleven models were evaluated, encompassing classical-based, transformer-based, and hybrid approaches. Using Precision, Recall, and F1-Score metrics, the evaluation indicated that hybrid approaches consistently outperform both classicalbased and standalone transformer-based approaches in legal entity extraction tasks.

Identifier

• OpenAlex: https://openalex.org/W7108332740
• DOI: https://doi.org/10.5753/eniac.2025.14243